{"id":90241,"date":"2026-09-30T17:47:36","date_gmt":"2026-09-30T09:47:36","guid":{"rendered":"https:\/\/wp-productionenv-bjg9h2g2bgg5b8aa.southeastasia-01.azurewebsites.net\/?p=90241"},"modified":"2026-09-30T17:59:55","modified_gmt":"2026-09-30T09:59:55","slug":"how-ai-detects-flood-changes-across-optical-and-sar-images-with-few-training-pairs","status":"publish","type":"post","link":"https:\/\/starpath.global\/blog\/how-ai-detects-flood-changes-across-optical-and-sar-images-with-few-training-pairs\/","title":{"rendered":"How AI Detects Flood Changes Across Optical and SAR Images With Few Training Pairs"},"content":{"rendered":"

In brief: <\/strong>Instead of separately mapping water in pre- and post-flood images and then overlaying the results, the method first maps water features from optical and SAR imagery into a more consistent semantic space, then directly identifies where water conditions have changed. More importantly, it does not require large numbers of real flood-change labels: it can automatically construct pseudo-change samples from existing water-body datasets.<\/p>\n

Flood emergency remote sensing faces a very practical dilemma.<\/p>\n

Before a flood, clear optical imagery is often available. Afterward, however, heavy rain and thick cloud cover may persist, making SAR imagery\u2014which can acquire data through clouds and in most weather conditions\u2014the more reliable option.<\/p>\n

As a result, the algorithm often receives not two visually similar images, but one optical image and one SAR image.<\/p>\n

That creates a fundamental problem.<\/p>\n

Optical imagery records reflected spectral information, while SAR measures microwave backscatter. The two look as though they speak entirely different languages. How can a change-detection model compare them directly?<\/strong><\/p>\n

In the paper Flood inundation monitoring using multi-source satellite imagery: a knowledge transfer strategy for heterogeneous image change detection<\/em>, published in Remote Sensing of Environment<\/em>, Wuhan University researchers Bofei Zhao, Haigang Sui and their colleagues proposed the Heterogeneous-Flood Inundation Extraction Network, or H-FIENet.<\/p>\n

The work does not simply introduce a deeper neural network. It addresses two persistent challenges in emergency flood mapping at the same time: the mismatch between features in heterogeneous imagery and the shortage of real, bi-temporal flood-change samples.<\/p>\n

1. Before Discussing the Network: Why Is \u201cPre-Flood Optical + Post-Flood SAR\u201d So Difficult?<\/h2>\n

\"The<\/p>\n

The Simplest Approach Is to Map Water Separately and Then Subtract the Results<\/h3>\n

A common method in conventional flood mapping is Post-Classification Comparison, or PCC.<\/p>\n

Its logic is straightforward:<\/p>\n

First, extract normal water bodies from the pre-flood image. Next, extract water from the post-flood image. Finally, compare the two binary water maps.<\/p>\n

If a location was not water before the flood but becomes water afterward, it is classified as flood expansion. If it was water before but is no longer water afterward, it represents flood recession.<\/p>\n

This sounds reasonable, but there is a major problem:<\/p>\n

The accuracy of the final flood map depends on both the pre-flood and post-flood water-extraction results being correct.<\/strong><\/p>\n

Any misclassification in either image propagates into the final change map.<\/p>\n

In SAR imagery, roads, building shadows and smooth land surfaces can all produce low backscatter and may therefore be mistaken for water. Errors from single-date water extraction are consequently carried into the flood map.<\/p>\n

Why Not Let Deep Learning Compare the Two Images Directly?<\/h3>\n

That is precisely what change-detection networks are designed to do.<\/p>\n

But this introduces a second problem: optical and SAR imagery are fundamentally different types of data.<\/strong><\/p>\n

Optical imagery depends on visible and near-infrared reflectance. SAR imagery depends on the interaction between microwaves and surface structure, roughness and dielectric properties.<\/p>\n

The same river may appear as a dark, spectrally stable area in an optical image, while in SAR imagery it is primarily characterized by low backscatter.<\/p>\n

The assumption underlying a conventional Siamese network\u2014that its two branches can share weights\u2014is therefore not naturally suited to this combination.<\/p>\n

The Harder Problem Is the Shortage of High-Quality Flood-Change Labels<\/h3>\n

If hundreds of thousands of precisely co-registered pre-flood optical images, post-flood SAR images and pixel-level flood-change labels were available, a model could gradually learn the necessary relationships.<\/p>\n

In reality, the opposite is true.<\/p>\n

Large water-body extraction datasets are widely available, but datasets containing pixel-level flood-change labels across different satellites, regions and acquisition dates remain extremely limited.<\/p>\n

Creating such samples manually is time-consuming, while flood emergencies themselves demand rapid response.<\/p>\n

The paper therefore addresses two problems:<\/p>\n

First, optical and SAR imagery occupy different feature spaces.<\/strong><\/p>\n

Second, heterogeneous flood-change samples are severely limited.<\/strong><\/p>\n

H-FIENet was developed around these two challenges.<\/p>\n

2. H-FIENet\u2019s Core Idea: Do Not Learn \u201cFlooding\u201d From Scratch\u2014Transfer Existing Knowledge About Water<\/h2>\n

\"The<\/p>\n

The authors identified an important fact:<\/p>\n

Although heterogeneous flood-change data are scarce, single-date water-body extraction data are not.<\/p>\n

Rather than asking a new model to learn from scratch what water is, what water looks like in optical imagery, what it looks like in SAR imagery and when a change should be classified as flooding, the method first inherits basic water knowledge from pretrained water-extraction models.<\/p>\n

The overall logic can be summarized in one sentence:<\/p>\n

First, teach the optical and SAR branches to recognize water independently. Then let the change-detection module learn what has happened to the water.<\/strong><\/p>\n

The paper calls this design cross-task knowledge transfer<\/strong>.<\/p>\n

Water-body extraction is the source task, while flood-change detection is the target task.<\/p>\n

H-FIENet therefore does not simply transfer a conventional classification model to a flood application. It transfers the water-related semantic knowledge most relevant to flood detection.<\/p>\n

3. The First Key Innovation: Optical and SAR Each Speak Their Own Language<\/h2>\n

\"The<\/p>\n

\"Feature<\/p>\n

A conventional Siamese network usually consists of two branches with identical structures and shared parameters.<\/p>\n

This is reasonable for change detection using homogeneous imagery because the pre- and post-event images come from the same or similar sensors and generally have consistent feature representations.<\/p>\n

For optical and SAR imagery, however, sharing parameters can become a constraint.<\/p>\n

H-FIENet instead adopts a pseudo-Siamese architecture.<\/p>\n

The two branches do not share parameters. Each uses a water-feature extraction network suited to its own data type.<\/p>\n

In the paper:<\/p>\n

    \n
  • The optical imagery branch uses MSResNet.<\/li>\n
  • The SAR imagery branch uses Siam-DWENet.<\/li>\n<\/ul>\n

    The two models are first trained separately on optical and SAR water-body datasets.<\/p>\n

    The authors then remove the output components originally used for single-date water classification while retaining the intermediate ability to extract water-related features. These become H-FIENet\u2019s two feature transformation modules.<\/p>\n

    The process therefore becomes:<\/p>\n

    \"Flowchart\"<\/p>\n

    What matters here is not simply that the network has two branches.<\/p>\n

    Instead of directly comparing raw optical and SAR pixels, the branches first transform their inputs into more comparable water-related semantic representations.<\/p>\n

    In other words, the model does not have to answer:<\/p>\n

    \u201cIs this SAR intensity value similar to that optical pixel value?\u201d<\/p>\n

    It instead asks:<\/p>\n

    \u201cAfter sensor-specific water-feature extraction, has the water state represented at this location changed?\u201d<\/p>\n

    This is the core mechanism H-FIENet uses to address the feature mismatch between heterogeneous images.<\/p>\n

    4. Why Is Feature Alignment More Sensible Than Direct Image Translation?<\/h2>\n

    \"Feature<\/p>\n

    One common approach to optical\u2013SAR change detection is to \u201ctranslate\u201d one image type into the other and then calculate the difference.<\/p>\n

    For example, a model might attempt to regress pre-flood optical features into a representation resembling post-flood SAR features before comparing them.<\/p>\n

    But this creates another problem.<\/p>\n

    Images contain extensive information about buildings, roads, farmland, vegetation, water and other land-cover types. Flood detection, however, is concerned primarily with whether the water state has changed.<\/p>\n

    H-FIENet therefore does not require the two modalities to become fully consistent across every type of land-cover feature. Instead, it focuses on aligning deep semantic features related to water.<\/p>\n

    Figure 12 in the paper compares the feature transformation results produced by AGSCC, CAAE and H-FIENet.<\/p>\n

    AGSCC can reveal prominent change areas through feature regression, but its superpixel processing introduces jagged boundaries and loses detail. CAAE relies on unchanged areas for unsupervised feature learning, making it prone to missed detections when changed areas account for a large proportion of the image.<\/p>\n

    H-FIENet instead concentrates on water-related semantics.<\/p>\n

    Its goal is therefore not:<\/p>\n

    To make optical imagery look like SAR imagery.<\/p>\n

    It is:<\/p>\n

    To make \u201cwater\u201d in optical imagery and \u201cwater\u201d in SAR imagery easier to compare within a deep feature space.<\/strong><\/p>\n

    That is the most important point for understanding H-FIENet.<\/p>\n

    5. The Second Key Innovation: Creating Change Labels Without Real Paired Flood Samples<\/h2>\n

    \"Spatially<\/p>\n

    \"Heterogeneous<\/p>\n

    Resolving the feature mismatch is not enough. Deep-learning-based change detection still typically requires large volumes of labeled bi-temporal data.<\/p>\n

    One of the paper\u2019s most instructive ideas is to reconsider a basic assumption:<\/p>\n

    Must the pre-event and post-event images used for change-detection training come from the same location?<\/p>\n

    The authors\u2019 answer is no.<\/p>\n

    What the Model Really Needs to Learn Is a Change Between Water and Non-Water<\/h3>\n

    Suppose an optical image from China already has a water mask.<\/p>\n

    A SAR image from another region also has a water mask.<\/p>\n

    Conventional change-detection methods would not pair them because they depict entirely different locations.<\/p>\n

    The authors argue, however, that geographic correspondence is unnecessary if the sole objective is to teach the model that a transition from water to non-water\u2014or from non-water to water\u2014represents change.<\/p>\n

    Optical water samples and SAR water samples from different regions can therefore be paired at random.<\/p>\n

    Only their water labels are then compared.<\/p>\n

    If the two labels indicate different states, the pixel is marked as \u201cchanged.\u201d If they indicate the same state, it is marked as \u201cunchanged.\u201d<\/p>\n

    The logic is straightforward:<\/p>\n

    \n\n\n\n\n\n\n\n\n
    Optical Water Label<\/th>\nSAR Water Label<\/th>\nPseudo-Change Label<\/th>\n<\/tr>\n<\/thead>\n
    Water<\/td>\nWater<\/td>\nUnchanged<\/td>\n<\/tr>\n
    Non-water<\/td>\nNon-water<\/td>\nUnchanged<\/td>\n<\/tr>\n
    Non-water<\/td>\nWater<\/td>\nChanged<\/td>\n<\/tr>\n
    Water<\/td>\nNon-water<\/td>\nChanged<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

    In essence, the method applies an exclusive-or operation to the two water masks.<\/p>\n

    The result is a pseudo flood dataset generated from spatially inconsistent multi-source pre- and post-event imagery.<\/p>\n

    Why Does This Matter?<\/h3>\n

    Because it changes how training samples can be obtained.<\/p>\n

    The conventional route is:<\/p>\n

    Find pre- and post-event images of the same location \u2192 precisely co-register them \u2192 manually label flood-related changes<\/p>\n

    H-FIENet instead uses:<\/p>\n

    Existing optical water data + existing SAR water data \u2192 random pairing \u2192 automatic generation of pseudo-change labels<\/p>\n

    Flood-change labels are scarce, but water masks are widely available.<\/p>\n

    The authors therefore turn a difficult data-acquisition problem into one that can reuse existing public datasets.<\/p>\n

    6. Can the Model Really Be Trained With Just Nine Sample Pairs? Transfer Learning\u2014not the Number Alone\u2014Makes It Possible<\/h2>\n

    One number in the paper immediately stands out:<\/p>\n

    H-FIENet uses only nine sets of 512 \u00d7 512 image patches for flood-change training.<\/p>\n

    Each set contains a pre-event optical image, a post-event SAR image and an automatically generated pseudo-flood-change label. The nine sets are divided among training, testing and validation at a ratio of 4:4:1.<\/p>\n

    A conventional deep network initialized with random weights would be unlikely to learn complex water representations across different satellite sensors reliably from only nine change-detection samples.<\/p>\n

    H-FIENet, however, does not start from scratch.<\/p>\n

    Its two front-end branches have already been pretrained on optical and SAR water datasets, allowing them to acquire substantial knowledge about water features.<\/p>\n

    The final few-shot training stage is therefore less about teaching the model what water is and more about teaching it how to compare water features that have already been identified.<\/p>\n

    This is why the paper\u2019s two main elements must be considered together:<\/p>\n

    Cross-task knowledge transfer + pseudo-change sample generation<\/p>\n

    Pseudo-samples alone would not be enough to support a robust model without stable, pretrained water features. Pretrained features alone would not eliminate the training-data bottleneck without a low-cost way to generate change labels.<\/p>\n

    Together, the two components form a complete strategy.<\/p>\n

    7. How Was the Method Tested? Across Different Countries and Satellite Combinations<\/h2>\n

    To show that H-FIENet was not effective only for one pair of Sentinel images, the authors evaluated it across multiple flood scenarios and satellite combinations.<\/p>\n

    The principal datasets included:<\/p>\n

    \n\n\n\n\n\n\n\n\n
    Data Source<\/th>\nType<\/th>\nRole in the Experiments<\/th>\n<\/tr>\n<\/thead>\n
    Sentinel-2<\/td>\nOptical<\/td>\nPre-flood background and water conditions<\/td>\n<\/tr>\n
    Sentinel-1<\/td>\nSAR<\/td>\nPost-flood observation<\/td>\n<\/tr>\n
    GF-2<\/td>\nOptical<\/td>\nTest of another high-resolution optical source<\/td>\n<\/tr>\n
    GF-3<\/td>\nSAR<\/td>\nTest of another SAR source<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

    The study areas included flood-affected regions around Wangjiaba in China, Li\u00e8ge in Belgium and Maputo in Mozambique.<\/p>\n

    These areas differ substantially in land cover, encompassing farmland, settlements, rivers, wetlands and vegetation across different climatic and geographic environments.<\/p>\n

    The experiments were therefore designed to answer more than whether the model could accurately map one flood.<\/p>\n

    The broader question was:<\/p>\n

    Would the feature-transfer strategy continue to work in another country, over different terrain or with a different satellite combination?<\/p>\n

    8. How Did It Perform? Overall Accuracy Reached 0.943 at Wangjiaba<\/h2>\n

    \"Flood<\/p>\n

    For the Wangjiaba experiment, the authors used pre-event Sentinel-2 optical imagery and post-event Sentinel-1 SAR imagery to identify flood inundation.<\/p>\n

    H-FIENet was compared with CMCDNet, GIR-MRF, AGSCC, CAAE and INLPG.<\/p>\n

    The main results were:<\/p>\n

    \n\n\n\n\n\n\n\n\n\n\n
    Method<\/th>\nOA<\/th>\nFA<\/th>\nMD<\/th>\nKappa<\/th>\n<\/tr>\n<\/thead>\n
    AGSCC<\/td>\n0.880<\/td>\n0.050<\/td>\n0.414<\/td>\n0.581<\/td>\n<\/tr>\n
    GIR-MRF<\/td>\n0.880<\/td>\n0.076<\/td>\n0.302<\/td>\n0.618<\/td>\n<\/tr>\n
    CMCDNet<\/td>\n0.930<\/td>\n0.040<\/td>\n0.175<\/td>\n0.787<\/td>\n<\/tr>\n
    CAAE<\/td>\n0.765<\/td>\n0.269<\/td>\n0.088<\/td>\n0.428<\/td>\n<\/tr>\n
    INLPG<\/td>\n0.744<\/td>\n0.188<\/td>\n0.541<\/td>\n0.256<\/td>\n<\/tr>\n
    H-FIENet<\/td>\n0.943<\/td>\n0.032<\/td>\n0.163<\/td>\n0.815<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

    The most important point is not simply that 0.943 is higher than 0.930.<\/p>\n

    CMCDNet is a supervised heterogeneous change-detection method that requires a larger number of bi-temporal training samples. H-FIENet achieved comparable or better results using only a small number of pseudo-change samples.<\/p>\n

    The paper also reports that H-FIENet processed the 3,800 \u00d7 3,100-pixel pre- and post-event images in approximately one minute.<\/p>\n

    That processing speed brings the method closer to the rapid-response requirements of emergency flood mapping.<\/p>\n

    9. Does It Still Work in Europe and Africa?<\/h2>\n

    \"Flood<\/p>\n

    \"Flood<\/p>\n

    The authors conducted further tests in Belgium and Mozambique, using flood products from the Copernicus Emergency Management Service as reference data.<\/p>\n

    H-FIENet achieved the following overall accuracy scores across the different scenarios:<\/p>\n